4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ptlrpc/pinger.c
38 * Portal-RPC reconnection and replay operations, for use in recovery.
41 #define DEBUG_SUBSYSTEM S_RPC
43 #include <obd_support.h>
44 #include <obd_class.h>
45 #include "ptlrpc_internal.h"
47 static int suppress_pings;
48 CFS_MODULE_PARM(suppress_pings, "i", int, 0644, "Suppress pings");
50 struct mutex pinger_mutex;
51 static struct list_head pinger_imports =
52 LIST_HEAD_INIT(pinger_imports);
53 static struct list_head timeout_list =
54 LIST_HEAD_INIT(timeout_list);
56 int ptlrpc_pinger_suppress_pings()
58 return suppress_pings;
60 EXPORT_SYMBOL(ptlrpc_pinger_suppress_pings);
62 struct ptlrpc_request *
63 ptlrpc_prep_ping(struct obd_import *imp)
65 struct ptlrpc_request *req;
67 req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING,
68 LUSTRE_OBD_VERSION, OBD_PING);
70 ptlrpc_request_set_replen(req);
71 req->rq_no_resend = req->rq_no_delay = 1;
76 int ptlrpc_obd_ping(struct obd_device *obd)
79 struct ptlrpc_request *req;
82 req = ptlrpc_prep_ping(obd->u.cli.cl_import);
86 req->rq_send_state = LUSTRE_IMP_FULL;
88 rc = ptlrpc_queue_wait(req);
90 ptlrpc_req_finished(req);
94 EXPORT_SYMBOL(ptlrpc_obd_ping);
96 static int ptlrpc_ping(struct obd_import *imp)
98 struct ptlrpc_request *req;
101 req = ptlrpc_prep_ping(imp);
103 CERROR("OOM trying to ping %s->%s\n",
104 imp->imp_obd->obd_uuid.uuid,
105 obd2cli_tgt(imp->imp_obd));
109 DEBUG_REQ(D_INFO, req, "pinging %s->%s",
110 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
111 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
116 static void ptlrpc_update_next_ping(struct obd_import *imp, int soon)
119 int time = soon ? PING_INTERVAL_SHORT : PING_INTERVAL;
120 if (imp->imp_state == LUSTRE_IMP_DISCON) {
121 int dtime = max_t(int, CONNECTION_SWITCH_MIN,
123 at_get(&imp->imp_at.iat_net_latency));
124 time = min(time, dtime);
126 imp->imp_next_ping = cfs_time_shift(time);
127 #endif /* ENABLE_PINGER */
130 void ptlrpc_ping_import_soon(struct obd_import *imp)
132 imp->imp_next_ping = cfs_time_current();
135 static inline int imp_is_deactive(struct obd_import *imp)
137 return (imp->imp_deactive ||
138 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_IMP_DEACTIVE));
141 static inline int ptlrpc_next_reconnect(struct obd_import *imp)
143 if (imp->imp_server_timeout)
144 return cfs_time_shift(obd_timeout / 2);
146 return cfs_time_shift(obd_timeout);
149 static cfs_duration_t pinger_check_timeout(cfs_time_t time)
151 struct timeout_item *item;
152 cfs_time_t timeout = PING_INTERVAL;
154 /* This list is sorted in increasing timeout order */
155 mutex_lock(&pinger_mutex);
156 list_for_each_entry(item, &timeout_list, ti_chain) {
157 int ti_timeout = item->ti_timeout;
158 if (timeout > ti_timeout)
159 timeout = ti_timeout;
162 mutex_unlock(&pinger_mutex);
164 return cfs_time_sub(cfs_time_add(time, cfs_time_seconds(timeout)),
171 void ptlrpc_pinger_ir_up(void)
173 CDEBUG(D_HA, "IR up\n");
176 EXPORT_SYMBOL(ptlrpc_pinger_ir_up);
178 void ptlrpc_pinger_ir_down(void)
180 CDEBUG(D_HA, "IR down\n");
183 EXPORT_SYMBOL(ptlrpc_pinger_ir_down);
185 static void ptlrpc_pinger_process_import(struct obd_import *imp,
186 unsigned long this_ping)
193 spin_lock(&imp->imp_lock);
195 level = imp->imp_state;
196 force = imp->imp_force_verify;
197 force_next = imp->imp_force_next_verify;
199 * This will be used below only if the import is "FULL".
201 suppress = ir_up && OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS);
203 imp->imp_force_verify = 0;
205 if (cfs_time_aftereq(imp->imp_next_ping - 5 * CFS_TICK, this_ping) &&
207 spin_unlock(&imp->imp_lock);
211 imp->imp_force_next_verify = 0;
213 spin_unlock(&imp->imp_lock);
215 CDEBUG(level == LUSTRE_IMP_FULL ? D_INFO : D_HA, "%s->%s: level %s/%u "
216 "force %u force_next %u deactive %u pingable %u suppress %u\n",
217 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
218 ptlrpc_import_state_name(level), level, force, force_next,
219 imp->imp_deactive, imp->imp_pingable, suppress);
221 if (level == LUSTRE_IMP_DISCON && !imp_is_deactive(imp)) {
222 /* wait for a while before trying recovery again */
223 imp->imp_next_ping = ptlrpc_next_reconnect(imp);
224 if (!imp->imp_no_pinger_recover)
225 ptlrpc_initiate_recovery(imp);
226 } else if (level != LUSTRE_IMP_FULL ||
227 imp->imp_obd->obd_no_recov ||
228 imp_is_deactive(imp)) {
229 CDEBUG(D_HA, "%s->%s: not pinging (in recovery "
230 "or recovery disabled: %s)\n",
231 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd),
232 ptlrpc_import_state_name(level));
234 spin_lock(&imp->imp_lock);
235 imp->imp_force_verify = 1;
236 spin_unlock(&imp->imp_lock);
238 } else if ((imp->imp_pingable && !suppress) || force_next || force) {
243 static int ptlrpc_pinger_main(void *arg)
245 struct ptlrpc_thread *thread = (struct ptlrpc_thread *)arg;
248 /* Record that the thread is running */
249 thread_set_flags(thread, SVC_RUNNING);
250 wake_up(&thread->t_ctl_waitq);
252 /* And now, loop forever, pinging as needed. */
254 cfs_time_t this_ping = cfs_time_current();
255 struct l_wait_info lwi;
256 cfs_duration_t time_to_next_wake;
257 struct timeout_item *item;
258 struct list_head *iter;
260 mutex_lock(&pinger_mutex);
261 list_for_each_entry(item, &timeout_list, ti_chain)
262 item->ti_cb(item, item->ti_cb_data);
264 list_for_each(iter, &pinger_imports) {
265 struct obd_import *imp = list_entry(iter,
269 ptlrpc_pinger_process_import(imp, this_ping);
270 /* obd_timeout might have changed */
271 if (imp->imp_pingable && imp->imp_next_ping &&
272 cfs_time_after(imp->imp_next_ping,
273 cfs_time_add(this_ping,
274 cfs_time_seconds(PING_INTERVAL))))
275 ptlrpc_update_next_ping(imp, 0);
277 mutex_unlock(&pinger_mutex);
278 /* update memory usage info */
279 obd_update_maxusage();
281 /* Wait until the next ping time, or until we're stopped. */
282 time_to_next_wake = pinger_check_timeout(this_ping);
283 /* The ping sent by ptlrpc_send_rpc may get sent out
284 say .01 second after this.
285 ptlrpc_pinger_sending_on_import will then set the
286 next ping time to next_ping + .01 sec, which means
287 we will SKIP the next ping at next_ping, and the
288 ping will get sent 2 timeouts from now! Beware. */
289 CDEBUG(D_INFO, "next wakeup in "CFS_DURATION_T" ("
290 CFS_TIME_T")\n", time_to_next_wake,
291 cfs_time_add(this_ping,cfs_time_seconds(PING_INTERVAL)));
292 if (time_to_next_wake > 0) {
293 lwi = LWI_TIMEOUT(max_t(cfs_duration_t,
295 cfs_time_seconds(1)),
297 l_wait_event(thread->t_ctl_waitq,
298 thread_is_stopping(thread) ||
299 thread_is_event(thread),
301 if (thread_test_and_clear_flags(thread, SVC_STOPPING)) {
305 /* woken after adding import to reset timer */
306 thread_test_and_clear_flags(thread, SVC_EVENT);
311 thread_set_flags(thread, SVC_STOPPED);
312 wake_up(&thread->t_ctl_waitq);
314 CDEBUG(D_NET, "pinger thread exiting, process %d\n", current_pid());
318 static struct ptlrpc_thread pinger_thread;
320 int ptlrpc_start_pinger(void)
322 struct l_wait_info lwi = { 0 };
323 struct task_struct *task;
325 #ifndef ENABLE_PINGER
330 if (!thread_is_init(&pinger_thread) &&
331 !thread_is_stopped(&pinger_thread))
334 init_waitqueue_head(&pinger_thread.t_ctl_waitq);
336 strcpy(pinger_thread.t_name, "ll_ping");
338 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
339 * just drop the VM and FILES in kthread_run() right away. */
340 task = kthread_run(ptlrpc_pinger_main, &pinger_thread,
341 pinger_thread.t_name);
344 CERROR("cannot start pinger thread: rc = %d\n", rc);
348 l_wait_event(pinger_thread.t_ctl_waitq,
349 thread_is_running(&pinger_thread), &lwi);
352 CWARN("Pings will be suppressed at the request of the "
353 "administrator. The configuration shall meet the "
354 "additional requirements described in the manual. "
355 "(Search for the \"suppress_pings\" kernel module "
361 int ptlrpc_pinger_remove_timeouts(void);
363 int ptlrpc_stop_pinger(void)
365 struct l_wait_info lwi = { 0 };
366 #ifndef ENABLE_PINGER
371 if (thread_is_init(&pinger_thread) ||
372 thread_is_stopped(&pinger_thread))
375 ptlrpc_pinger_remove_timeouts();
377 thread_set_flags(&pinger_thread, SVC_STOPPING);
378 wake_up(&pinger_thread.t_ctl_waitq);
380 l_wait_event(pinger_thread.t_ctl_waitq,
381 thread_is_stopped(&pinger_thread), &lwi);
385 void ptlrpc_pinger_sending_on_import(struct obd_import *imp)
387 ptlrpc_update_next_ping(imp, 0);
390 void ptlrpc_pinger_commit_expected(struct obd_import *imp)
392 ptlrpc_update_next_ping(imp, 1);
393 assert_spin_locked(&imp->imp_lock);
395 * Avoid reading stale imp_connect_data. When not sure if pings are
396 * expected or not on next connection, we assume they are not and force
397 * one anyway to guarantee the chance of updating
398 * imp_peer_committed_transno.
400 if (imp->imp_state != LUSTRE_IMP_FULL ||
401 OCD_HAS_FLAG(&imp->imp_connect_data, PINGLESS))
402 imp->imp_force_next_verify = 1;
405 int ptlrpc_pinger_add_import(struct obd_import *imp)
408 if (!list_empty(&imp->imp_pinger_chain))
411 mutex_lock(&pinger_mutex);
412 CDEBUG(D_HA, "adding pingable import %s->%s\n",
413 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
414 /* if we add to pinger we want recovery on this import */
415 imp->imp_obd->obd_no_recov = 0;
416 ptlrpc_update_next_ping(imp, 0);
417 /* XXX sort, blah blah */
418 list_add_tail(&imp->imp_pinger_chain, &pinger_imports);
419 class_import_get(imp);
421 ptlrpc_pinger_wake_up();
422 mutex_unlock(&pinger_mutex);
426 EXPORT_SYMBOL(ptlrpc_pinger_add_import);
428 int ptlrpc_pinger_del_import(struct obd_import *imp)
432 if (list_empty(&imp->imp_pinger_chain))
435 mutex_lock(&pinger_mutex);
436 list_del_init(&imp->imp_pinger_chain);
437 CDEBUG(D_HA, "removing pingable import %s->%s\n",
438 imp->imp_obd->obd_uuid.uuid, obd2cli_tgt(imp->imp_obd));
439 /* if we remove from pinger we don't want recovery on this import */
440 imp->imp_obd->obd_no_recov = 1;
441 class_import_put(imp);
442 mutex_unlock(&pinger_mutex);
445 EXPORT_SYMBOL(ptlrpc_pinger_del_import);
448 * Register a timeout callback to the pinger list, and the callback will
449 * be called when timeout happens.
451 static struct timeout_item *ptlrpc_new_timeout(int time,
452 enum timeout_event event,
453 timeout_cb_t cb, void *data)
455 struct timeout_item *ti;
461 INIT_LIST_HEAD(&ti->ti_obd_list);
462 INIT_LIST_HEAD(&ti->ti_chain);
463 ti->ti_timeout = time;
464 ti->ti_event = event;
466 ti->ti_cb_data = data;
472 * Register timeout event on the the pinger thread.
473 * Note: the timeout list is an sorted list with increased timeout value.
475 static struct timeout_item*
476 ptlrpc_pinger_register_timeout(int time, enum timeout_event event,
477 timeout_cb_t cb, void *data)
479 struct timeout_item *item, *tmp;
481 LASSERT(mutex_is_locked(&pinger_mutex));
483 list_for_each_entry(item, &timeout_list, ti_chain)
484 if (item->ti_event == event)
487 item = ptlrpc_new_timeout(time, event, cb, data);
489 list_for_each_entry_reverse(tmp, &timeout_list, ti_chain) {
490 if (tmp->ti_timeout < time) {
491 list_add(&item->ti_chain, &tmp->ti_chain);
495 list_add(&item->ti_chain, &timeout_list);
501 /* Add a client_obd to the timeout event list, when timeout(@time)
502 * happens, the callback(@cb) will be called.
504 int ptlrpc_add_timeout_client(int time, enum timeout_event event,
505 timeout_cb_t cb, void *data,
506 struct list_head *obd_list)
508 struct timeout_item *ti;
510 mutex_lock(&pinger_mutex);
511 ti = ptlrpc_pinger_register_timeout(time, event, cb, data);
513 mutex_unlock(&pinger_mutex);
516 list_add(obd_list, &ti->ti_obd_list);
517 mutex_unlock(&pinger_mutex);
520 EXPORT_SYMBOL(ptlrpc_add_timeout_client);
522 int ptlrpc_del_timeout_client(struct list_head *obd_list,
523 enum timeout_event event)
525 struct timeout_item *ti = NULL, *item;
527 if (list_empty(obd_list))
529 mutex_lock(&pinger_mutex);
530 list_del_init(obd_list);
532 * If there are no obd attached to the timeout event
533 * list, remove this timeout event from the pinger
535 list_for_each_entry(item, &timeout_list, ti_chain) {
536 if (item->ti_event == event) {
541 LASSERTF(ti != NULL, "ti is NULL !\n");
542 if (list_empty(&ti->ti_obd_list)) {
543 list_del(&ti->ti_chain);
546 mutex_unlock(&pinger_mutex);
549 EXPORT_SYMBOL(ptlrpc_del_timeout_client);
551 int ptlrpc_pinger_remove_timeouts(void)
553 struct timeout_item *item, *tmp;
555 mutex_lock(&pinger_mutex);
556 list_for_each_entry_safe(item, tmp, &timeout_list, ti_chain) {
557 LASSERT(list_empty(&item->ti_obd_list));
558 list_del(&item->ti_chain);
561 mutex_unlock(&pinger_mutex);
565 void ptlrpc_pinger_wake_up()
568 thread_add_flags(&pinger_thread, SVC_EVENT);
569 wake_up(&pinger_thread.t_ctl_waitq);
573 /* Ping evictor thread */
575 #define PET_TERMINATE 2
577 static int pet_refcount = 0;
578 static int pet_state;
579 static wait_queue_head_t pet_waitq;
580 static struct list_head pet_list;
581 static DEFINE_SPINLOCK(pet_lock);
583 int ping_evictor_wake(struct obd_export *exp)
585 struct obd_device *obd;
587 spin_lock(&pet_lock);
588 if (pet_state != PET_READY) {
589 /* eventually the new obd will call here again. */
590 spin_unlock(&pet_lock);
594 obd = class_exp2obd(exp);
595 if (list_empty(&obd->obd_evict_list)) {
596 class_incref(obd, "evictor", obd);
597 list_add(&obd->obd_evict_list, &pet_list);
599 spin_unlock(&pet_lock);
605 static int ping_evictor_main(void *arg)
607 struct obd_device *obd;
608 struct obd_export *exp;
609 struct l_wait_info lwi = { 0 };
615 CDEBUG(D_HA, "Starting Ping Evictor\n");
616 pet_state = PET_READY;
618 l_wait_event(pet_waitq, (!list_empty(&pet_list)) ||
619 (pet_state == PET_TERMINATE), &lwi);
621 /* loop until all obd's will be removed */
622 if ((pet_state == PET_TERMINATE) && list_empty(&pet_list))
625 /* we only get here if pet_exp != NULL, and the end of this
626 * loop is the only place which sets it NULL again, so lock
627 * is not strictly necessary. */
628 spin_lock(&pet_lock);
629 obd = list_entry(pet_list.next, struct obd_device,
631 spin_unlock(&pet_lock);
633 expire_time = cfs_time_current_sec() - PING_EVICT_TIMEOUT;
635 CDEBUG(D_HA, "evicting all exports of obd %s older than %ld\n",
636 obd->obd_name, expire_time);
638 /* Exports can't be deleted out of the list while we hold
639 * the obd lock (class_unlink_export), which means we can't
640 * lose the last ref on the export. If they've already been
641 * removed from the list, we won't find them here. */
642 spin_lock(&obd->obd_dev_lock);
643 while (!list_empty(&obd->obd_exports_timed)) {
644 exp = list_entry(obd->obd_exports_timed.next,
646 exp_obd_chain_timed);
647 if (expire_time > exp->exp_last_request_time) {
648 class_export_get(exp);
649 spin_unlock(&obd->obd_dev_lock);
650 LCONSOLE_WARN("%s: haven't heard from client %s"
651 " (at %s) in %ld seconds. I think"
652 " it's dead, and I am evicting"
653 " it. exp %p, cur %ld expire %ld"
656 obd_uuid2str(&exp->exp_client_uuid),
657 obd_export_nid2str(exp),
658 (long)(cfs_time_current_sec() -
659 exp->exp_last_request_time),
660 exp, (long)cfs_time_current_sec(),
662 (long)exp->exp_last_request_time);
663 CDEBUG(D_HA, "Last request was at %ld\n",
664 exp->exp_last_request_time);
665 class_fail_export(exp);
666 class_export_put(exp);
667 spin_lock(&obd->obd_dev_lock);
669 /* List is sorted, so everyone below is ok */
673 spin_unlock(&obd->obd_dev_lock);
675 spin_lock(&pet_lock);
676 list_del_init(&obd->obd_evict_list);
677 spin_unlock(&pet_lock);
679 class_decref(obd, "evictor", obd);
681 CDEBUG(D_HA, "Exiting Ping Evictor\n");
686 void ping_evictor_start(void)
688 struct task_struct *task;
690 if (++pet_refcount > 1)
693 INIT_LIST_HEAD(&pet_list);
694 init_waitqueue_head(&pet_waitq);
696 task = kthread_run(ping_evictor_main, NULL, "ll_evictor");
699 CERROR("Cannot start ping evictor thread: %ld\n",
703 EXPORT_SYMBOL(ping_evictor_start);
705 void ping_evictor_stop(void)
707 if (--pet_refcount > 0)
710 pet_state = PET_TERMINATE;
713 EXPORT_SYMBOL(ping_evictor_stop);